Every time someone types a product review, answers a question on a forum, or fires off a comment at the bottom of an article, they are doing something they almost certainly do not think about.
They are making a commodity, one that a rapidly expanding industry now depends on to function.
And researchers tracking that industry just put a number on the whole thing that is hard to ignore.
The invisible raw material hiding in plain sight every day
Think about the sheer volume of text produced online in a single afternoon.
Product reviews, Reddit threads, recipe comments, customer service transcripts, news article replies, billions of words, typed by ordinary people, uploaded and largely forgotten.
For decades, that content sat online as background noise.
It was useful for search engines, sure. But nobody thought of it as an industrial raw material with a measurable market price.
That assumption is now completely out of date.
A gold rush that most people never noticed was happening
To train a large language model, the kind that powers modern AI assistants, you need text.
Enormous quantities of it, covering as many subjects, styles and sentence structures as possible.
The most powerful AI systems to date are language models trained primarily on trillions of words of human-generated text from the internet, according to Epoch AI.
That research compared the current situation to a finite natural resource being depleted at pace, a dynamic researchers have likened to a gold rush in a finite mine.
And once an industry starts paying for a raw material, the material stops being free.
Much like the hidden value economists calculated for generative AI tools used by millions of Americans, the true worth of what ordinary people produce online is only now being properly measured.
The deals that revealed what a typed sentence is actually worth
The clearest sign that human text has a market price is the licensing deal.
Reddit signed a content licensing deal worth about $60 million per year with Google to train its Gemini AI models, an agreement reached in early 2024, just before Reddit’s IPO filing.
That is $60 million a year for the right to use words that everyday users typed for free, often just to share an opinion or ask for advice.
In its IPO prospectus, Reddit disclosed data licensing arrangements with an aggregate contract value of $203 million and terms of two to three years.
These deals are not outliers. They map a market that barely existed three years ago: the global AI training dataset market was valued at $3.59 billion in 2025 and is projected to grow to $4.44 billion in 2026, according to Fortune Business Insights.
The $4.44 billion market built entirely on words people wrote for free
Here is the reveal the numbers have been building toward.
The global AI training dataset market is projected to grow from $4.44 billion in 2026 to $23.18 billion by 2034, at a compound annual growth rate of 22.90%.
The raw material powering that market is human-written text, the very thing that anyone with a keyboard has been producing and giving away for free since the early days of the web.
But here is the twist that researchers say should concern the whole industry: the effective stock of quality and repetition-adjusted human-generated public text for AI training is estimated at around 300 trillion tokens, and if trends continue, language models will fully utilize this stock between 2026 and 2032, or even earlier if intensely overtrained.
The total stock of publicly available human-generated text is large but bounded, and frontier models consume data faster than the internet produces it at current growth rates.
A market worth billions of dollars is pressing against the limits of its own supply, and the limiting factor is simply the pace at which human beings write things down.
This scarcity dynamic is already reshaping how organizations from universities to the military think about AI infrastructure investment.
What happens when the words run out, and what comes next
When public data eventually runs out, developers will have to decide what to feed language models.
Ideas include data now considered private, like emails or text messages, and using synthetic data created by other AI models.
Synthetic data generation is already growing rapidly as a way to address data scarcity challenges in niche applications, and researchers see it as one of the most promising paths forward.
There is something quietly remarkable about all of this.
The text that people have been typing casually online for the past 30 years, reviews, arguments, explanations, jokes, turns out to have been among the most valuable datasets ever assembled.
It was collected not by scientists or engineers, but by ordinary people sharing ordinary thoughts, one sentence at a time.
Researchers and companies are now working hard on synthetic alternatives, and the industry is not standing still.
But for now, the most irreplaceable ingredient in a multi-billion-dollar market remains the one thing no model has yet learned to replace: a human, sitting down, and deciding to write something real.
